Lecture 1 - What is AI? Turing test and microworlds

• 1st try: AI is the study of intelligent systems (not the opposite of natural intelligence)
• But that would include anthropology!
• 2nd try: study of how to build of intelligent systems
• What is intelligence?
• If we knew the problem would be much easier
• 1st try: given situation do the right thing
• right thing implies a performance or cost function
• this would make AI merely a particular case of optimization theory
• it doesn't turn out that way since the problems are hard
• Is a thrown ball intelligent? Parabolic path is the result of an extremely complex computation!
• 2nd try: everything that has not yet done by computer
• too explicitly time dependent
• 3rd try: model intelligence on the intelligent systems we know (e.g. humans)
• Turing test
• Does intelligence imply consciousness?
• Can we prove consciousness from input-output relations?
• Can I prove that any other human is conscious?
• Solipsism (Zen, Descartes, Idealism, disconnected brain conjecture, transformed disconnected brain)
• Deterministic chaos - even simple machines can perform unexpectably
• Chinese room thought experiment
• strong vs. weak AI
• If full Turing test is the target then no-one is working on AI yet !
• Microworlds - the way out
• well defined and manageable sub-problems of the complete Turing test
• similar to models in science - not intended to be full solution
• Examples:
• Winograd's SHRDLU blocks world
• Game playing (sheshbesh, chess, checkers)
• Speech and speaker recognition
• Optical character recognition (OCR)
• Automatic translation (blind idiot, vodka is strong but meat is rotten)
• Expert systems
• Artificial life (ALIFE)
• Simple example Tic Tac Toe
• finite but big for lookup table - 9! = 362880 different boards, so < 3^9 different games
• solvable game (first step - take middle square, etc.)
• Solution 1: hand wired
• Solution 2: look-up table 3^9 entries, do what is stored (hard but FAST)
• Solution 3: simple expert system heuristics
• A first take middle; B win if you can; C block if you can; D random move
• Solution 4: look-ahead
• investigate all possible moves and use the best
• which is best? one way: evaluate recursively! (minimax criterion)
• which is best? another way: use stored heuristic function
• chess programs use both recursive look-ahead and board cost heuristics
• Solution 5: solution 2 or 4 with learning.
• play against humans or against other programs
• Solution 6: full learning function
• parametric, build architecture as you go
• feedforward neural network approach (9-X-9 representation)
• Summary: What is intelligence?
• Is best performer necessarily intelligent?
• GOFAI (Good Old Fashioned AI - van Neuman architecture can do everything)
• vs NN (neural network) - where there is an existence proof!